When the cost of producing code approaches zero, AI-generated code that 'looks plausible but contains subtle defects' is often harder to detect than obvious errors in human-written code.